NVIDIA Introduces Kumo Relational, a Foundation Model That Predicts Directly Over Enterprise Databases
The Apache 2.0 model claims to beat feature-engineered XGBoost pipelines on 41 tasks across 15 databases without any per-task training.
NVIDIA announced Kumo Relational on 21 October at GTC Berlin, a pre-trained relational foundation model that makes predictions such as churn, demand forecasting, fraud and imputation directly on top of relational databases, without feature engineering or per-task training. It applies in-context learning to relational data: up to 10,000 historical examples are read as context in a single forward pass. A Relational Graph Transformer architecture processes tables as-is, using primary keys, foreign keys and timestamps without flattening.
NVIDIA claims it is the first few-shot model to beat supervised ML on 41 tasks across 15 databases including RelBench V1/V2, SAP SALT and 4DBInfer, surpassing feature-engineered XGBoost pipelines by about 4 AUROC points on RelBench V1, with scale to over 500 billion rows via a memory-mapped graph engine running up to 5GB/s. The library is Apache 2.0, available as a Python package and as an NIM container. Performance figures come from NVIDIA and NVIDIA-backed research.